chore: import upstream snapshot with attribution

2026-07-13 12:44:17 +08:00
commit 85742ab165
588 changed files with 320176 additions and 0 deletions
@@ -0,0 +1,32 @@
+name: Backport Merged Pull Request
+on:
+  pull_request_target:
+    types: [closed]
+permissions:
+  contents: write
+  issues: write
+  pull-requests: write
+
+# NOTE:
+# Microsoft requires rotating BOT_PAT every 3 months.
+# Log onto agent-lightning-bot account and rotate the PAT if needed.
+
+jobs:
+  backport:
+    name: Backport pull request
+    runs-on: ubuntu-latest
+    # Don't run on closed unmerged pull requests
+    if: github.event.pull_request.merged
+    steps:
+      - uses: actions/checkout@v6
+      - name: Create backport pull requests
+        uses: korthout/backport-action@v3
+        with:
+          branch_name: 'backport/${pull_number}/${target_branch}'
+          label_pattern: ^(stable/[^ ]+)$
+          github_token: ${{ secrets.BOT_PAT }}
+          add_labels: backport
+          add_author_as_assignee: true
+          git_committer_name: agent-lightning-bot
+          # This email address is not monitored.
+          git_committer_email: agl.msft@outlook.com
@@ -0,0 +1,29 @@
+name: Badge - APO
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - APO
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-apo.yml', label: 'apo', variants: ['legacy', 'stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - Azure
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Azure
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-azure.yml', label: 'azure', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - Calc-X
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Calc-X
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-calc-x.yml', label: 'calc-x', variants: ['legacy', 'stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - ChartQA
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - ChartQA
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-chartqa.yml', label: 'chartqa', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - Claude Code
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Claude Code
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-claude-code.yml', label: 'claude-code', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - Compatibility
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Backward Compatibility
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-compat.yml', label: 'examples-compat', variants: ['legacy', 'stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,45 @@
+name: Badge - Examples
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Calc-X
+      - Examples - Spider
+      - Examples - APO
+      - Examples - Unsloth
+      - Examples - Tinker
+      - Examples - Azure
+      - Examples - Claude Code
+      - Examples - RAG
+      - Examples - ChartQA
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-calc-x.yml', label: 'examples-calc-x.stable', variants: ['stable'] },
+              { workflow: 'examples-spider.yml', label: 'examples-spider.stable', variants: ['stable'] },
+              { workflow: 'examples-apo.yml', label: 'examples-apo.stable', variants: ['stable'] },
+              { workflow: 'examples-unsloth.yml', label: 'examples-unsloth.stable', variants: ['stable'] },
+              { workflow: 'examples-tinker.yml', label: 'examples-tinker.stable', variants: ['stable'] },
+              { workflow: 'examples-azure.yml', label: 'examples-azure.stable', variants: ['stable'] },
+              { workflow: 'examples-claude-code.yml', label: 'examples-claude-code.stable', variants: ['stable'] },
+              { workflow: 'examples-rag.yml', label: 'examples-rag.stable', variants: ['stable'] },
+              { workflow: 'examples-chartqa.yml', label: 'examples-chartqa.stable', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,41 @@
+name: Badge - Latest
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Calc-X
+      - Examples - Spider
+      - Examples - APO
+      - Examples - Unsloth
+      - Examples - RAG
+      - Examples - Claude Code
+      - GPU Test
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-calc-x.yml', label: 'calc-x.latest', variants: ['latest'] },
+              { workflow: 'examples-spider.yml', label: 'spider.latest', variants: ['latest'] },
+              { workflow: 'examples-apo.yml', label: 'apo.latest', variants: ['latest'] },
+              { workflow: 'examples-unsloth.yml', label: 'unsloth.latest', variants: ['latest'] },
+              { workflow: 'examples-claude-code.yml', label: 'claude-code.latest', variants: ['latest'] },
+              { workflow: 'examples-rag.yml', label: 'rag.latest', variants: ['latest'] },
+              { workflow: 'tests-full.yml', label: 'tests-full.latest', variants: ['latest'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - RAG
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - RAG
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-rag.yml', label: 'rag', variants: ['legacy', 'stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - Spider
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Spider
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-spider.yml', label: 'spider', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - Tinker
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Tinker
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-tinker.yml', label: 'tinker', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,31 @@
+name: Badge - Unit Test
+
+on:
+  workflow_run:
+    workflows:
+      - CPU Test
+      - GPU Test
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'tests-full.yml', label: 'tests-full', variants: ['legacy', 'stable'] },
+              { workflow: 'tests.yml', label: 'tests', variants: ['legacy', 'stable', 'Lint', 'documentation', 'JavaScript'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,29 @@
+name: Badge - Unsloth
+
+on:
+  workflow_run:
+    workflows:
+      - Examples - Unsloth
+    types: [completed]
+
+  workflow_dispatch:
+
+permissions:
+  actions: read
+  contents: read
+
+jobs:
+  badge:
+    if: ${{ github.event_name == 'workflow_dispatch' || (github.event_name == 'workflow_run' && github.event.workflow_run.head_branch == 'main') }}
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/github-script@v8
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          script: |
+            const badgeAggregation = require('./scripts/badge_aggregation.js');
+            const dependencies = [
+              { workflow: 'examples-unsloth.yml', label: 'examples-unsloth.stable', variants: ['stable'] },
+            ];
+            await badgeAggregation({ github, context, core, dependencies });
@@ -0,0 +1,441 @@
+name: Benchmark
+permissions:
+  contents: read
+on:
+  workflow_dispatch:
+  schedule:
+    # Every Monday and Thursday at 3 AM UTC+8
+    - cron: '0 19 * * 0,3'
+
+jobs:
+  benchmark:
+    name: ${{ matrix.workload.kind }} (${{ matrix.backend.id }}, ${{ matrix.workload.display }})
+    runs-on: ${{ matrix.workload.runner }}
+    timeout-minutes: ${{ matrix.workload.timeout }}
+    strategy:
+      fail-fast: false
+      matrix:
+        backend:
+          - id: memory
+            compose_file: compose.prometheus-memory-store.yml
+          - id: mongo
+            compose_file: compose.prometheus-mongo-store.yml
+        workload:
+          - id: scenario-minimal-scale
+            display: Minimal production scale
+            kind: scenario
+            store_workers: 4
+            runner:
+              - self-hosted
+              - 1ES.Pool=agl-runner-cpu
+            timeout: 45
+            args: >-
+              --mode batch
+              --total-tasks 4096
+              --batch-size 256
+              --n-runners 32
+              --max-rounds 6
+              --sleep-seconds 0.5
+          - id: scenario-medium-scale
+            display: Medium production scale
+            kind: scenario
+            store_workers: 16
+            runner:
+              - self-hosted
+              - 1ES.Pool=agl-runner-cpu
+            timeout: 45
+            args: >-
+              --mode batch
+              --total-tasks 10000
+              --batch-size 1000
+              --n-runners 100
+              --max-rounds 10
+              --sleep-seconds 0.1
+          - id: scenario-midhigh-scale
+            display: Mid-high production scale
+            kind: scenario
+            store_workers: 24
+            runner:
+              - self-hosted
+              - 1ES.Pool=agl-runner-cpu
+            timeout: 60
+            args: >-
+              --mode batch
+              --total-tasks 20000
+              --batch-size 2048
+              --n-runners 300
+              --max-rounds 6
+              --sleep-seconds 0.1
+          - id: scenario-large-batch
+            display: Large batch waves
+            kind: scenario
+            store_workers: 96
+            runner:
+              - self-hosted
+              - 1ES.Pool=agl-runner-cpu-high
+            timeout: 120
+            args: >-
+              --mode batch
+              --total-tasks 50000
+              --batch-size 8192
+              --n-runners 1000
+              --max-rounds 3
+              --sleep-seconds 0.1
+          - id: scenario-long-queues
+            display: Long rollout queues
+            kind: scenario
+            store_workers: 48
+            runner:
+              - self-hosted
+              - 1ES.Pool=agl-runner-cpu
+            timeout: 120
+            args: >-
+              --mode batch_partial
+              --total-tasks 50000
+              --batch-size 1024
+              --n-runners 256
+              --remaining-tasks 4096
+              --max-rounds 4
+              --sleep-seconds 0.1
+          - id: scenario-high-concurrency
+            display: High-throughput concurrent requests
+            kind: scenario
+            store_workers: 96
+            runner:
+              - self-hosted
+              - 1ES.Pool=agl-runner-cpu-high
+            timeout: 120
+            args: >-
+              --mode single
+              --total-tasks 50000
+              --concurrency 2048
+              --n-runners 256
+              --max-rounds 2
+              --sleep-seconds 0.1
+          - id: scenario-heavy-traces
+            display: Heavy rollouts with deep traces
+            kind: scenario
+            store_workers: 96
+            runner:
+              - self-hosted
+              - 1ES.Pool=agl-runner-cpu-high
+            timeout: 60
+            args: >-
+              --mode batch_partial
+              --total-tasks 10000
+              --batch-size 1024
+              --remaining-tasks 256
+              --n-runners 512
+              --max-rounds 20
+              --sleep-seconds 1.0
+
+          - id: micro-worker
+            display: Update worker
+            kind: micro
+            store_workers: 8
+            runner: ubuntu-latest
+            timeout: 30
+            cli: worker
+          - id: micro-dequeue-empty
+            display: Dequeue empty
+            kind: micro
+            store_workers: 8
+            runner: ubuntu-latest
+            timeout: 30
+            cli: dequeue-empty
+          - id: micro-rollout
+            display: Rollout + span
+            kind: micro
+            store_workers: 8
+            runner: ubuntu-latest
+            timeout: 30
+            cli: rollout
+          - id: micro-dequeue-update-attempt
+            display: Dequeue + update attempt
+            kind: micro
+            store_workers: 8
+            runner: ubuntu-latest
+            timeout: 30
+            cli: dequeue-update-attempt
+          - id: micro-dequeue-only
+            display: Dequeue only
+            kind: micro
+            store_workers: 8
+            runner: ubuntu-latest
+            timeout: 30
+            cli: dequeue-only
+          - id: micro-metrics
+            display: Multi-metric fan-out
+            kind: micro
+            store_workers: 8
+            runner: ubuntu-latest
+            timeout: 15
+            cli: metrics
+    env:
+      PYTHONUNBUFFERED: "1"
+      STORE_URL: http://localhost:4747
+      STORE_API_URL: http://localhost:4747/v1/agl
+      PROM_URL: http://localhost:9090
+      GITHUB_ACTIONS_TIMEOUT_MINUTES: ${{ matrix.workload.timeout }}
+      WORKLOAD_KIND: ${{ matrix.workload.kind }}
+      WORKLOAD_ID: ${{ matrix.workload.id }}
+      BACKEND_ID: ${{ matrix.backend.id }}
+      ARTIFACT_DIR: ${{ format('artifacts/{0}-{1}', matrix.workload.id, matrix.backend.id) }}
+      COMPOSE_FILE: ${{ matrix.backend.compose_file }}
+      AGL_STORE_N_WORKERS: ${{ matrix.workload.store_workers }}
+      ANALYSIS_FILE: ${{ format('analysis-{0}.log', matrix.workload.id) }}
+      SUMMARY_FILE: ${{ format('summary-{0}.log', matrix.workload.id) }}
+      PROM_ARCHIVE_BASENAME: ${{ format('prometheus-{0}-{1}', matrix.workload.id, matrix.backend.id) }}
+      ARTIFACT_NAME: ${{ format('{0}-{1}', matrix.workload.id, matrix.backend.id) }}
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: '3.12'
+
+      - name: Sync dependencies
+        run: uv sync --frozen --extra mongo --group core-stable --group dev
+
+      - name: Check disk space
+        run: df -h
+
+      - name: Reset benchmark data directories
+        run: |
+          set -euo pipefail
+          cd docker
+          rm -rf data
+          bash setup.sh
+
+      - name: Launch ${{ matrix.backend.id }} Prometheus stack
+        run: |
+          set -euo pipefail
+          cd docker
+          docker compose -f "$COMPOSE_FILE" down -v || true
+          docker compose -f "$COMPOSE_FILE" up -d --quiet-pull
+
+      - name: Wait for store readiness
+        run: |
+          set -euo pipefail
+          for attempt in {1..60}; do
+            if curl -fsS "$STORE_API_URL/health" >/dev/null 2>&1; then
+              sleep 1
+              curl -fsS "$STORE_API_URL/rollouts"  # Warm up the scraper
+              sleep 15  # Allow some time for the baseline metrics to be established
+              exit 0
+            fi
+            sleep 1
+          done
+          echo "Store did not become ready in time" >&2
+          # show logs for debugging
+          cd docker && docker compose -f "$COMPOSE_FILE" logs app
+          exit 1
+
+      - name: Prepare artifact directory
+        run: mkdir -p "$ARTIFACT_DIR"
+
+      - name: Record workload start
+        run: echo "BENCHMARK_START=$(date -u +%FT%TZ)" >> "$GITHUB_ENV"
+
+      - name: (Scenario) Run ${{ matrix.workload.display }} workload
+        if: ${{ matrix.workload.kind == 'scenario' }}
+        run: |
+          set -euo pipefail
+          uv run --locked --no-sync python -m tests.benchmark.benchmark_store \
+            --store-url "$STORE_URL" \
+            ${{ matrix.workload.args }}
+
+      - name: (Micro) Run ${{ matrix.workload.display }}
+        if: ${{ matrix.workload.kind == 'micro' }}
+        run: |
+          set -euo pipefail
+          mkdir -p "$ARTIFACT_DIR"
+          uv run --locked --no-sync python -m tests.benchmark.micro_benchmark \
+            --store-url "$STORE_URL" \
+            --summary-file "$ARTIFACT_DIR/$SUMMARY_FILE" \
+            "${{ matrix.workload.cli }}" | tee "$ARTIFACT_DIR/${{ matrix.workload.id }}.txt"
+
+      - name: Record workload end
+        if: ${{ always() }}
+        run: echo "BENCHMARK_END=$(date -u +%FT%TZ)" >> "$GITHUB_ENV"
+
+      - name: Show micro benchmark summary
+        if: ${{ always() && matrix.workload.kind == 'micro' }}
+        run: |
+          set -euo pipefail
+          summary_file="$ARTIFACT_DIR/$SUMMARY_FILE"
+          if [ -f "$summary_file" ]; then
+            echo "Micro benchmark summary ($WORKLOAD_ID/$BACKEND_ID):"
+            cat "$summary_file"
+          else
+            echo "Summary file not found: $summary_file"
+          fi
+
+      - name: Run workload analysis
+        if: ${{ always() }}
+        run: |
+          set -euo pipefail
+          mkdir -p "$ARTIFACT_DIR"
+          if [ -z "${BENCHMARK_START:-}" ] || [ -z "${BENCHMARK_END:-}" ]; then
+            echo "Analysis skipped: benchmark window not recorded." > "$ARTIFACT_DIR/$ANALYSIS_FILE"
+            exit 1
+          fi
+          uv run --locked --no-sync python -m tests.benchmark.analysis \
+            --prom-url "$PROM_URL" \
+            --store-url "$STORE_API_URL" \
+            --start "$BENCHMARK_START" \
+            --end "$BENCHMARK_END" \
+            | tee "$ARTIFACT_DIR/$ANALYSIS_FILE"
+
+      - name: Collect docker logs
+        if: ${{ always() }}
+        run: |
+          set -euo pipefail
+          mkdir -p "$ARTIFACT_DIR"
+          cd docker
+          readarray -t services < <(docker compose -f "$COMPOSE_FILE" config --services)
+          if [ "${#services[@]}" -eq 0 ]; then
+            echo "No services defined in compose file."
+            exit 0
+          fi
+          for service in "${services[@]}"; do
+            docker compose -f "$COMPOSE_FILE" logs "$service" > "../$ARTIFACT_DIR/docker-${service}-${WORKLOAD_ID}-${BACKEND_ID}.log" || true
+          done
+
+      - name: Stop ${{ matrix.backend.id }} Prometheus stack
+        if: ${{ always() }}
+        run: |
+          set -euo pipefail
+          cd docker
+          docker compose -f "$COMPOSE_FILE" down -v || true
+
+      - name: Archive Prometheus metrics
+        if: ${{ always() }}
+        run: |
+          set -euo pipefail
+          mkdir -p "$ARTIFACT_DIR"
+          if [ -d docker/data/prometheus ]; then
+            tar -C docker/data -czf "$ARTIFACT_DIR/${PROM_ARCHIVE_BASENAME}.tar.gz" prometheus
+          fi
+
+      - name: Upload workload artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v6
+        with:
+          name: ${{ env.ARTIFACT_NAME }}
+          path: ${{ env.ARTIFACT_DIR }}
+          if-no-files-found: error
+
+  collection-benchmarks:
+    name: collection (${{ matrix.backend.id }}, ${{ matrix.workload.id }})
+    runs-on: ${{ matrix.backend.runner }}
+    timeout-minutes: 15
+    strategy:
+      fail-fast: false
+      matrix:
+        backend:
+          - id: memory
+            needs_mongo: false
+            runner: ubuntu-latest
+          - id: mongo
+            needs_mongo: true
+            runner: ubuntu-latest
+        workload:
+          - id: high-insert
+            total_tasks: 50000
+            concurrency: 2048
+            type: insert
+          - id: medium-insert
+            total_tasks: 50000
+            concurrency: 128
+            type: insert
+          - id: low-insert
+            total_tasks: 50000
+            concurrency: 4
+            type: insert
+          - id: high-dequeue
+            total_tasks: 50000
+            concurrency: 2048
+            type: dequeue
+          - id: medium-dequeue
+            total_tasks: 50000
+            concurrency: 128
+            type: dequeue
+          - id: low-dequeue
+            total_tasks: 50000
+            concurrency: 4
+            type: dequeue
+    env:
+      ARTIFACT_DIR: ${{ format('artifacts/{0}-{1}', matrix.backend.id, matrix.workload.id) }}
+      SUMMARY_FILE: ${{ format('artifacts/{0}-{1}/summary-{0}-{1}.jsonl', matrix.backend.id, matrix.workload.id) }}
+      ARTIFACT_NAME: ${{ format('collections-{0}-{1}', matrix.backend.id, matrix.workload.id) }}
+      MONGO_URI: mongodb://localhost:27017/?replicaSet=rs0
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: '3.12'
+
+      - name: Sync dependencies
+        run: uv sync --frozen --extra mongo --group core-stable --group dev
+
+      - name: Launch MongoDB
+        if: ${{ matrix.backend.needs_mongo }}
+        run: |
+          set -euo pipefail
+          cd docker
+          docker compose -f compose.mongo.yml down -v || true
+          docker compose -f compose.mongo.yml up -d --quiet-pull
+          for attempt in {1..60}; do
+            if docker compose -f compose.mongo.yml exec -T mongo mongosh --quiet --eval 'db.runCommand({ping:1})' >/dev/null 2>&1; then
+              exit 0
+            fi
+            sleep 2
+          done
+          echo "MongoDB did not become ready in time" >&2
+          docker compose -f compose.mongo.yml logs mongo
+          exit 1
+
+      - name: Run collection benchmark
+        run: |
+          set -euo pipefail
+          mkdir -p "$ARTIFACT_DIR"
+          echo "Running collection benchmark (backend=${{ matrix.backend.id }}, workload=${{ matrix.workload.id }})"
+          uv run --locked --no-sync python -m tests.benchmark.collection_benchmark \
+            "${{ matrix.workload.type }}" \
+            --backend "${{ matrix.backend.id }}" \
+            --total-tasks "${{ matrix.workload.total_tasks }}" \
+            --concurrency "${{ matrix.workload.concurrency }}" \
+            --task-prefix "${{ matrix.backend.id }}-${{ matrix.workload.id }}" \
+            --summary-file "$SUMMARY_FILE" \
+            --mongo-uri "$MONGO_URI" \
+            --mongo-database agentlightning_collection_bench
+
+      - name: Show collection benchmark summary
+        if: ${{ always() }}
+        run: |
+          set -euo pipefail
+          if [ -f "$SUMMARY_FILE" ]; then
+            echo "Collection benchmark summary (${{ matrix.backend.id }}):"
+            cat "$SUMMARY_FILE"
+          else
+            echo "Summary file not found: $SUMMARY_FILE"
+          fi
+
+      - name: Stop MongoDB
+        if: ${{ always() && matrix.backend.needs_mongo }}
+        run: |
+          set -euo pipefail
+          cd docker
+          docker compose -f compose.mongo.yml down -v || true
+
+      - name: Upload collection artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v6
+        with:
+          name: ${{ env.ARTIFACT_NAME }}
+          path: ${{ env.ARTIFACT_DIR }}
+          if-no-files-found: error
@@ -0,0 +1,33 @@
+name: Dashboard
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 5 AM UTC+8
+    - cron: '0 21 * * *'
+
+  workflow_dispatch:
+
+  push:
+    branches: [ main, stable/**/* ]
+
+jobs:
+  dashboard:
+    name: Chromatic
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Run Chromatic
+        uses: chromaui/action@v13
+        with:
+          projectToken: ${{ secrets.CHROMATIC_PROJECT_TOKEN }}
+          workingDir: dashboard
+          exitZeroOnChanges: false
@@ -0,0 +1,64 @@
+name: Deploy Documentation
+
+on:
+  push:
+    branches:
+      - main
+    tags:
+      - 'v*'
+  workflow_dispatch:
+
+concurrency:
+  group: docs-deploy
+  cancel-in-progress: false
+
+permissions:
+  contents: write
+  pages: write
+  id-token: write
+
+jobs:
+  deploy:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@v6
+        with:
+          python-version: '3.12'
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+      - name: Sync dependencies
+        run: uv sync --frozen --no-default-groups --group dev
+
+      - name: Configure Git
+        run: |
+          git config --global user.name "GitHub Actions"
+          git config --global user.email "actions@github.com"
+
+      - name: Get version and commit
+        id: version
+        run: |
+          if [[ $GITHUB_REF == refs/tags/* ]]; then
+            VERSION=${GITHUB_REF#refs/tags/v}
+            SOURCE_COMMIT=${GITHUB_SHA}
+          else
+            VERSION=$(python -c "import tomllib; print(tomllib.load(open('pyproject.toml', 'rb'))['project']['version'])")
+            SOURCE_COMMIT=${GITHUB_SHA}
+          fi
+          echo "version=$VERSION" >> $GITHUB_OUTPUT
+          echo "SOURCE_COMMIT=$SOURCE_COMMIT" >> $GITHUB_ENV
+
+      - name: Deploy versioned docs
+        if: startsWith(github.ref, 'refs/tags/')
+        run: |
+          uv run --locked --no-sync mike deploy --push --update-aliases ${{ steps.version.outputs.version }} stable
+
+      - name: Deploy dev docs
+        if: github.ref == 'refs/heads/main'
+        run: |
+          uv run --locked --no-sync mike deploy --push latest
+          # Always set stable to default
+          uv run --locked --no-sync mike set-default --push stable
@@ -0,0 +1,116 @@
+name: Examples - APO
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 3 AM UTC+8
+    - cron: '0 19 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-apo, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'APO - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('APO - {0}', github.event_name) }}
+
+jobs:
+  apo:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-apo' ||
+      github.event.action == 'ci-all'
+    name: APO (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    # This job is run on GitHub hosted runners rather than self-hosted runners because it needs no GPU.
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: |
+          uv sync --frozen --no-default-groups --extra apo \
+            --group dev --group experiment --group agents --group core-stable
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (stable & legacy)
+        run: |
+          uv sync --frozen --no-default-groups --extra apo \
+            --group dev --group experiment --group agents --group core-${{ matrix.setup-script }}
+        if: matrix.setup-script != 'latest'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-apo-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: APO custom algorithm
+        run: |
+          set -ex
+          cd examples/apo
+          uv run apo_custom_algorithm_trainer.py | tee _ci_apo.log
+          # Check whether the log contains "Best prompt found:"
+          grep "Best prompt found:" _ci_apo.log
+        env:
+          # New versions follow OPENAI_BASE_URL instead of OPENAI_API_BASE
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+      - name: APO custom algorithm debugger
+        run: |
+          set -ex
+          cd examples/apo
+          uv run apo_debug.py --mode runner
+          uv run apo_debug.py --mode hook
+          uv run apo_debug.py --mode trainer
+        env:
+          # New versions follow OPENAI_BASE_URL instead of OPENAI_API_BASE
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+
+      - name: APO built-in algorithm
+        run: |
+          set -ex
+          cd examples/apo
+          uv run room_selector_apo.py
+        env:
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+        if: matrix.setup-script != 'legacy'
@@ -0,0 +1,98 @@
+name: Examples - Azure
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 4 AM UTC+8
+    - cron: '0 20 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-azure, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Azure - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Azure - {0}', github.event_name) }}
+
+jobs:
+  azure:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-azure' ||
+      github.event.action == 'ci-all'
+    name: Azure (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-cpu]
+    timeout-minutes: 400
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.12'
+          setup-script: 'stable'
+      fail-fast: false
+    steps:
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups \
+            --group dev --group experiment --group agents --group core-stable
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-azure-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Azure Login
+        run: |
+          az login --identity
+        shell: bash
+
+      - name: Azure OpenAI Sanity Check
+        run: |
+          source .venv/bin/activate
+          cd examples/azure
+          python capital_agent.py
+        shell: bash
+        env:
+          AZURE_OPENAI_ENDPOINT: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          AZURE_OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+        id: azure_openai_sanity_check
+
+      - name: Azure OpenAI Supervised Fine-tuning
+        run: |
+          source .venv/bin/activate
+          cd examples/azure
+          python train_capital_agent.py --n-iterations 2 --cleanup
+        shell: bash
+        env:
+          AZURE_OPENAI_ENDPOINT: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          AZURE_OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          AZURE_SUBSCRIPTION_ID: ${{ secrets.AZURE_SUBSCRIPTION_ID }}
+          AZURE_OPENAI_API_VERSION: 2025-04-01-preview
+          AZURE_RESOURCE_GROUP: ${{ secrets.AZURE_RESOURCE_GROUP }}
+          AZURE_RESOURCE_NAME: ${{ secrets.AZURE_RESOURCE_NAME }}
+        id: azure_openai_finetune
@@ -0,0 +1,438 @@
+name: Examples - Calc-X
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 3 AM UTC+8
+    - cron: '0 19 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-calc-x, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Calc-X - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Calc-X - {0}', github.event_name) }}
+
+jobs:
+  calc-x-perf:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-calc-x' ||
+      github.event.action == 'ci-all'
+    name: Calc-X Performance (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 90
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --group torch-gpu-stable
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (stable & legacy)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --group torch-gpu-${{ matrix.setup-script }}
+        if: matrix.setup-script != 'latest'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-calc-x-performance-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: Prepare Calc-X dataset
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run gdown --fuzzy https://drive.google.com/file/d/1FQMyKLLd6hP9dw9rfZn1EZOWNvKaDsqw/view
+          unzip calc-x-data.zip -d data
+          rm calc-x-data.zip
+
+      - name: Calc-X MCP sanity check
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run tests/test_mcp_calculator.py
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+      - name: Calc-X sanity check
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run legacy_calc_agent_debug.py
+        env:
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+
+      # Calc-X training suddenly works after running the sanity check.
+      # And it has to be run before Spider training.
+      # The client side used to hang in many of my attempts.
+      # Don't ask why. Don't touch this.
+      - name: Calc-X training
+        run: |
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          python train_calc_agent.py --val-file data/test_mini.parquet --ci
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train
+
+      - name: Validate Calc-X training
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train.outputs.project_name }} ${{ steps.calc_x_train.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+  calc-x-variants:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-calc-x' ||
+      github.event.action == 'ci-all'
+    name: Calc-X Variants (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 90
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --extra weave --extra mongo --group torch-gpu-stable
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (stable & legacy)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --extra weave --extra mongo --group torch-gpu-${{ matrix.setup-script }}
+        if: matrix.setup-script != 'latest'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-calc-x-variants-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: Prepare Calc-X dataset
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run gdown --fuzzy https://drive.google.com/file/d/1FQMyKLLd6hP9dw9rfZn1EZOWNvKaDsqw/view
+          unzip calc-x-data.zip -d data
+          rm calc-x-data.zip
+
+      - name: Calc-X MCP sanity check
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run tests/test_mcp_calculator.py
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+      - name: Calc-X sanity check
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run legacy_calc_agent_debug.py
+        env:
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+
+      - name: Training with local model
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          hf download Qwen/Qwen2.5-0.5B-Instruct --local-dir data/qwen_model
+          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast --model $(realpath data/qwen_model)
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_local_model
+
+      - name: Validate training with local model
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_local_model.outputs.project_name }} ${{ steps.calc_x_train_local_model.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Training with LLM Proxy
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast --llm-proxy
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_llm_proxy
+
+      - name: Validate training with LLM Proxy
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_llm_proxy.outputs.project_name }} ${{ steps.calc_x_train_llm_proxy.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Setup Docker environments
+        run: ./scripts/mongodb_docker_run.sh
+        shell: bash
+
+      - name: Training with MongoDB
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast --mongo-uri mongodb://localhost:27017/?replicaSet=rs0
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_mongo
+
+      - name: Validate training with MongoDB
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_mongo.outputs.project_name }} ${{ steps.calc_x_train_mongo.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Training with LoRA
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast --lora
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_lora
+        if: matrix.setup-script != 'legacy'
+
+      - name: Validate training with LoRA
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_lora.outputs.project_name }} ${{ steps.calc_x_train_lora.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        if: matrix.setup-script != 'legacy'
+
+      - name: Training with trajectory level aggregation
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast --trajectory-level
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_trajectory_level
+
+      - name: Validate training with trajectory level aggregation
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_trajectory_level.outputs.project_name }} ${{ steps.calc_x_train_trajectory_level.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Training with Weave
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast --weave
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_weave
+
+      - name: Validate training with Weave
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_weave.outputs.project_name }} ${{ steps.calc_x_train_weave.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Training with external store
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+
+          agl store --port 4747 &
+          sleep 5
+          AGL_MANAGED_STORE=0 AGL_CURRENT_ROLE=runner python train_calc_agent.py --external-store-address http://localhost:4747 --val-file data/test_mini.parquet --ci-fast &
+          sleep 5
+          AGL_MANAGED_STORE=0 AGL_CURRENT_ROLE=algorithm python train_calc_agent.py --external-store-address http://localhost:4747 --val-file data/test_mini.parquet --ci-fast
+
+          pkill -f agl && echo "SIGTERM sent to agl" || echo "No agl process found"
+          while pgrep -f agl; do
+            echo "Waiting for agl to finish..."
+            sleep 5
+          done
+          pkill -f train_calc_agent.py && echo "SIGTERM sent to train_calc_agent.py" || echo "No train_calc_agent.py process found"
+          while pgrep -f train_calc_agent.py; do
+            echo "Waiting for train_calc_agent.py to finish..."
+            sleep 5
+          done
+          echo "train_calc_agent.py has finished."
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_external_store
+
+      - name: Validate training with external store
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_external_store.outputs.project_name }} ${{ steps.calc_x_train_external_store.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Training with role-based environment variables
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+
+          PYTHONUNBUFFERED=1 AGL_SERVER_HOST=127.0.0.1 AGL_SERVER_PORT=5858 AGL_CURRENT_ROLE=runner python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast &
+          sleep 5
+          PYTHONUNBUFFERED=1 AGL_SERVER_HOST=0.0.0.0 AGL_SERVER_PORT=5858 AGL_CURRENT_ROLE=algorithm python train_calc_agent.py --val-file data/test_mini.parquet --ci-fast
+
+          pkill -f train_calc_agent.py && echo "SIGTERM sent to train_calc_agent.py" || echo "No train_calc_agent.py process found"
+          while pgrep -f train_calc_agent.py; do
+            echo "Waiting for train_calc_agent.py to finish..."
+            sleep 5
+          done
+          echo "train_calc_agent.py has finished."
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train_role_based_env_var
+
+      - name: Validate training with role-based environment variables
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train_role_based_env_var.outputs.project_name }} ${{ steps.calc_x_train_role_based_env_var.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
@@ -0,0 +1,168 @@
+name: Examples - ChartQA
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 6 AM UTC+8
+    - cron: "0 22 * * *"
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-chartqa, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'ChartQA - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('ChartQA - {0}', github.event_name) }}
+
+jobs:
+  chartqa:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-chartqa' ||
+      github.event.action == 'ci-all'
+    name: ChartQA (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 60
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.12'
+          setup-script: 'stable'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group image --group langchain --group vllm-0-10-2 --group torch-gpu-stable
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-chartqa-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: Prepare ChartQA dataset
+        run: |
+          set -euo pipefail
+          cd examples/chartqa
+          uv run gdown --fuzzy "https://drive.google.com/file/d/1fWRt9hehg8_uV7BDWSCwKTycM60JcmGN/view?usp=sharing" -O chartqa-data.zip
+          unzip chartqa-data.zip
+          rm chartqa-data.zip
+        shell: bash
+
+      - name: ChartQA sanity check with GPT
+        run: |
+          set -euo pipefail
+          cd examples/chartqa
+          uv run python debug_chartqa_agent.py
+        shell: bash
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+
+      - name: Run vLLM Server
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/chartqa
+          uv run --no-sync vllm serve Qwen/Qwen2-VL-2B-Instruct \
+            --gpu-memory-utilization 0.9 \
+            --max-model-len 4096 \
+            --allowed-local-media-path "$(pwd)/data" \
+            --enable-prefix-caching \
+            --port 8088 &
+
+          VLLM_READY=0
+          for i in {1..100}; do
+            if curl -sSf http://localhost:8088/v1/models > /dev/null 2>&1; then
+              echo "vLLM server is ready!"
+              VLLM_READY=1
+              break
+            fi
+            echo "Waiting for vLLM server to be ready... (${i})"
+            sleep 5
+          done
+          if [[ "$VLLM_READY" != "1" ]]; then
+            echo "vLLM server failed to start!"
+            exit 1
+          fi
+
+      - name: ChartQA sanity check with vLLM
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/chartqa
+          uv run python debug_chartqa_agent.py
+        shell: bash
+        env:
+          USE_LLM_PROXY: "1"
+          OPENAI_API_BASE: http://localhost:8088/v1
+          OPENAI_MODEL: Qwen/Qwen2-VL-2B-Instruct
+
+      - name: Stop vLLM Server
+        run: |
+          set -euo pipefail
+          pkill -f vllm
+          for i in {1..60}; do
+            if ! pgrep -f vllm; then
+              break
+            fi
+            sleep 5
+          done
+
+      - name: ChartQA training
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/chartqa
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_chartqa_agent.py ci
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: chartqa_train
+
+      - name: Validate ChartQA training
+        run: |
+          set -euo pipefail
+          uv run scripts/validate_example_wandb.py ${{ steps.chartqa_train.outputs.project_name }} ${{ steps.chartqa_train.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
@@ -0,0 +1,151 @@
+name: Examples - Claude Code
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 4 AM UTC+8
+    - cron: "0 20 * * *"
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-claude-code, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Claude Code - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Claude Code - {0}', github.event_name) }}
+
+jobs:
+  claude-code:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-claude-code' ||
+      github.event.action == 'ci-all'
+    name: Claude Code (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 60
+    strategy:
+      matrix:
+        include:
+          - python-version: "3.12"
+            setup-script: "stable"
+          - python-version: "3.13"
+            setup-script: "latest"
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups \
+            --group dev --group experiment --group agents --group torch-gpu-stable
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-claude-code-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Download model
+        run: |
+          source .venv/bin/activate
+          python -c "from transformers import AutoModelForCausalLM; AutoModelForCausalLM.from_pretrained('Qwen/Qwen3-Coder-30B-A3B-Instruct')"
+
+      - name: Launch vLLM server
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          vllm serve Qwen/Qwen3-Coder-30B-A3B-Instruct \
+            --max-model-len 131072 \
+            --enable-auto-tool-choice \
+            --tool-call-parser qwen3_coder \
+            --port 45993 &
+
+          VLLM_READY=0
+          for i in {1..100}; do
+            if curl -sSf http://localhost:45993/v1/models > /dev/null 2>&1; then
+              echo "vLLM server is ready!"
+              VLLM_READY=1
+              break
+            fi
+            echo "Waiting for vLLM server to be ready... (${i})"
+            sleep 5
+          done
+          if [[ "$VLLM_READY" != "1" ]]; then
+            echo "vLLM server failed to start!"
+            exit 1
+          fi
+
+      - name: Claude Code sanity check with vLLM models
+        run: |
+          source .venv/bin/activate
+          cd examples/claude_code
+          python claude_code_agent.py vllm --backend-model-high Qwen/Qwen3-Coder-30B-A3B-Instruct --backend-model-low Qwen/Qwen3-Coder-30B-A3B-Instruct --base-url http://localhost:45993/v1 --debug
+        shell: bash
+
+      - name: Upload sanity check artifacts for vLLM
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v6
+        with:
+          name: claude-code-sanity-check-vllm-${{ matrix.setup-script }}
+          path: |
+            examples/claude_code/data/
+            examples/claude_code/logs/
+          if-no-files-found: error
+
+      - name: Cleanup vLLM
+        run: |
+          set -euo pipefail
+          pkill -f vllm
+          for i in {1..60}; do
+            if ! pgrep -f vllm; then
+              break
+            fi
+            sleep 5
+          done
+          rm -rf examples/claude_code/data/
+          rm -rf examples/claude_code/logs/
+
+      - name: Claude Code sanity check with OpenAI models
+        run: |
+          source .venv/bin/activate
+          cd examples/claude_code
+          python claude_code_agent.py openai --backend-model-high gpt-5.1-codex-mini --backend-model-low gpt-4.1-mini --debug
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+
+      - name: Upload sanity check artifacts for OpenAI
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v6
+        with:
+          name: claude-code-sanity-check-openai-${{ matrix.setup-script }}
+          path: |
+            examples/claude_code/data/
+            examples/claude_code/logs/
+          if-no-files-found: error
@@ -0,0 +1,151 @@
+name: Examples - Backward Compatibility
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 6 AM UTC+8
+    - cron: '0 22 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-compat, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Backward Compatibility - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Backward Compatibility - {0}', github.event_name) }}
+
+jobs:
+  backward-compatibility:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-compat' ||
+      github.event.action == 'ci-all'
+    name: Backward Compatibility (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 30
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups --extra apo --extra verl \
+            --group dev --group experiment --group agents --group torch-gpu-${{ matrix.setup-script }}
+      - name: Override VERL (stable)
+        run: |
+          uv pip install verl==0.5.0 vllm==0.10.2
+        if: matrix.setup-script == 'stable'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-backward-compatibility-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+      - name: Prepare Calc-X dataset
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run gdown --fuzzy https://drive.google.com/file/d/1FQMyKLLd6hP9dw9rfZn1EZOWNvKaDsqw/view
+          unzip calc-x-data.zip -d data
+          rm calc-x-data.zip
+
+      - name: APO example (legacy client-server style)
+        run: |
+          set -ex
+          cd examples/apo
+          uv run legacy_apo_client.py &
+          sleep 3  # Wait for the client to be up
+          uv run legacy_apo_server.py
+          pkill -f legacy_apo_client.py && echo "SIGTERM sent to legacy_apo_client.py" || echo "No legacy_apo_client.py process found"
+          while pgrep -f legacy_apo_client.py; do
+            echo "Waiting for legacy_apo_client.py to finish..."
+            sleep 5
+          done
+          echo "legacy_apo_client.py has finished."
+          sleep 10
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+
+      - name: Calc-X MCP sanity check
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run tests/test_mcp_calculator.py
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+      - name: Calc-X sanity check
+        run: |
+          set -ex
+          cd examples/calc_x
+          uv run legacy_calc_agent_debug.py
+        env:
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+
+      - name: Calc-X training (legacy client-server style)
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/calc_x
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python legacy_calc_agent.py &
+          bash legacy_train.sh
+          pkill -f legacy_calc_agent.py && echo "SIGTERM sent to legacy_calc_agent.py" || echo "No legacy_calc_agent.py process found"
+          while pgrep -f legacy_calc_agent.py; do
+            echo "Waiting for legacy_calc_agent.py to finish..."
+            sleep 5
+          done
+          echo "legacy_calc_agent.py has finished."
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: calc_x_train
+
+      - name: Validate Calc-X training
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.calc_x_train.outputs.project_name }} ${{ steps.calc_x_train.outputs.run_name }}
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
@@ -0,0 +1,179 @@
+name: Examples - RAG
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 6 AM UTC+8
+    - cron: '0 22 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-rag, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'RAG - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('RAG - {0}', github.event_name) }}
+
+jobs:
+  rag:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-rag' ||
+      github.event.action == 'ci-all'
+    name: RAG (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 60
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --group rag --group torch-gpu-stable
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (stable & legacy)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --group rag --group torch-gpu-${{ matrix.setup-script }}
+        if: matrix.setup-script != 'latest'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-rag-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: Prepare RAG dataset
+        run: |
+          set -euo pipefail
+          cd examples/rag
+          mkdir -p data
+          uv run gdown --fuzzy "https://drive.google.com/file/d/1Pq4Ag8zVoN8gUtLu0LcBfY35Dm5zL0hq/view?usp=drive_link" -O data/dataset_tiny.parquet
+          uv run gdown --fuzzy "https://drive.google.com/file/d/1REXCpRLbeZu1KfWWKhIGEQe_WNHUOBkS/view?usp=drive_link" -O data/chunks_candidate_tiny.pkl
+          uv run gdown --fuzzy "https://drive.google.com/file/d/1f6P-h_8KSRhe5pqDHWbRQWvUhTygfZ-c/view?usp=drive_link" -O data/index_hnsw_faiss_n32e40_tiny.index
+
+      - name: Run WIKI Retriever MCP Server
+        run: |
+          set -euo pipefail
+          cd examples/rag
+          uv run python wiki_retriever_mcp.py &
+          for i in {1..20}; do
+            sleep 5
+            if nc -z localhost 8099; then
+              echo "MCP server is up!"
+              exit 0
+            else
+              echo "Waiting for MCP server to start..."
+            fi
+          done
+          echo "MCP server failed to start within expected time."
+          exit 1
+
+      - name: Run vLLM Server
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          vllm serve Qwen/Qwen2.5-1.5B-Instruct \
+            --enable-auto-tool-choice \
+            --tool-call-parser hermes \
+            --port 8000 &
+
+          VLLM_READY=0
+          for i in {1..100}; do
+            if curl -sSf http://localhost:8000/v1/models > /dev/null 2>&1; then
+              echo "vLLM server is ready!"
+              VLLM_READY=1
+              break
+            fi
+            echo "Waiting for vLLM server to be ready... (${i})"
+            sleep 5
+          done
+          if [[ "$VLLM_READY" != "1" ]]; then
+            echo "vLLM server failed to start!"
+            exit 1
+          fi
+
+      - name: Run RAG Sanity check
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/rag
+          uv run python rag_agent.py
+        shell: bash
+
+      - name: Stop vLLM Server
+        run: |
+          set -euo pipefail
+          pkill -f vllm
+          for i in {1..60}; do
+            if ! pgrep -f vllm; then
+              break
+            fi
+            sleep 5
+          done
+
+      - name: RAG training
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/rag
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_rag.py fast
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: rag_train
+
+      - name: Validate RAG training
+        run: |
+          set -ex
+          # Allow up to 5 rollouts to fail to produce rewards
+          uv run scripts/validate_example_wandb.py ${{ steps.rag_train.outputs.project_name }} ${{ steps.rag_train.outputs.run_name }} --reward-tolerance 5
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
@@ -0,0 +1,126 @@
+name: Examples - Spider
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 4 AM UTC+8
+    - cron: '0 20 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-spider, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Spider - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Spider - {0}', github.event_name) }}
+
+jobs:
+  spider:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-spider' ||
+      github.event.action == 'ci-all'
+    name: Spider (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 60
+    strategy:
+      matrix:
+        include:
+        # legacy is omitted because langchain doesn't work with legacy vllm versions
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --group langchain --group torch-gpu-stable
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (stable)
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group agents --group langchain --group torch-gpu-${{ matrix.setup-script }}
+        if: matrix.setup-script == 'stable'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-spider-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: Prepare Spider dataset
+        run: |
+          set -ex
+          cd examples/spider
+          uv run gdown --fuzzy https://drive.google.com/file/d/1oi9J1jZP9TyM35L85CL3qeGWl2jqlnL6/view
+          unzip -q spider-data.zip -d data
+          rm spider-data.zip
+
+      - name: Spider sanity check
+        run: |
+          set -ex
+          cd examples/spider
+          uv run sql_agent.py
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+        if: success() || failure()
+
+      - name: Spider training
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/spider
+          ../../scripts/restart_ray.sh
+          sleep 5
+          PYTHONUNBUFFERED=1 python train_sql_agent.py fast
+          sleep 10
+        shell: bash
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+        id: spider_train
+
+      - name: Validate Spider training
+        run: |
+          set -ex
+          uv run scripts/validate_example_wandb.py ${{ steps.spider_train.outputs.project_name }} ${{ steps.spider_train.outputs.run_name }} --reward-tolerance 5
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
@@ -0,0 +1,170 @@
+name: Examples - Tinker
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 3 AM UTC+8
+    - cron: '0 19 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-tinker, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Tinker - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Tinker - {0}', github.event_name) }}
+
+jobs:
+  tinker:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-tinker' ||
+      github.event.action == 'ci-all'
+    name: Tinker (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-cpu]
+    timeout-minutes: 150
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups \
+            --group dev --group experiment --group agents --group torch-cpu --group core-stable --group tinker
+      - name: Freeze dependencies
+        run: |
+          set -euo pipefail
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-tinker-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      # TODO: Currently only test the client tracer implementation.
+      - name: Tinker LLM sanity check (tracer text)
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          python -m tests.test_tinker_llm tracer-text
+        shell: bash
+        env:
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker LLM sanity check (tracer tool)
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          python -m tests.test_tinker_llm tracer-tool
+        shell: bash
+        env:
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Hello
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          python hello.py oneclick --ci
+        shell: bash
+        env:
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Q20 Evaluate (GPT-4.1)
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          mkdir -p logs
+          python q20_evaluate.py --ci --model gpt-4.1 --output-file logs/q20_evaluate_gpt-4.1.jsonl
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          CREWAI_DISABLE_TELEMETRY: true
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Q20 Evaluate (Qwen3-30B-A3B-Instruct-2507)
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          python q20_evaluate.py --ci --model Qwen/Qwen3-30B-A3B-Instruct-2507 --output-file logs/q20_evaluate_qwen3-30b-a3b.jsonl
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          CREWAI_DISABLE_TELEMETRY: true
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Q20 Training Dry Run
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          python q20_train.py dryrun --model qwen4b
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          CREWAI_DISABLE_TELEMETRY: true
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
+
+      - name: Tinker Q20 Training
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/tinker
+          agl store --port 4747 &
+          sleep 5
+          python q20_train.py runner --n-runners 4 &
+          sleep 5
+          python q20_train.py algo --model qwen4b --ci
+          sleep 5
+
+          pkill -f agl && echo "SIGTERM sent to agl" || echo "No agl process found"
+          while pgrep -f agl; do
+            echo "Waiting for agl to finish..."
+            sleep 5
+          done
+          pkill -f q20_train.py && echo "SIGTERM sent to q20_train.py" || echo "No q20_train.py process found"
+          while pgrep -f q20_train.py; do
+            echo "Waiting for q20_train.py to finish..."
+            sleep 5
+          done
+          echo "q20_train.py has finished."
+        shell: bash
+        env:
+          OPENAI_BASE_URL: ${{ secrets.AZURE_OPENAI_ENDPOINT_SWEDEN }}
+          OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY_SWEDEN }}
+          CREWAI_DISABLE_TELEMETRY: true
+          TINKER_API_KEY: ${{ secrets.TINKER_API_KEY }}
@@ -0,0 +1,129 @@
+name: Examples - Unsloth
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 5 AM UTC+8
+    - cron: '0 21 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-unsloth, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'Unsloth - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('Unsloth - {0}', github.event_name) }}
+
+jobs:
+  unsloth:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-unsloth' ||
+      github.event.action == 'ci-all'
+    name: Unsloth (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 60
+    strategy:
+      matrix:
+        # Legacy versions are not supported for Unsloth examples.
+        include:
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - name: Check disk space
+        run: df -h
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies
+        run: |
+          uv sync --frozen --no-default-groups --extra verl \
+            --group dev --group experiment --group trl --group agents --group torch-gpu-stable
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-unsloth-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Prepare Unsloth model
+        run: |
+          set -ex
+          cd examples/unsloth
+          rm -rf models
+          uv run hf download unsloth/Qwen3-4B-Instruct-2507 --local-dir models/version_0
+
+      - name: Unsloth SFT example
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/unsloth
+
+          agl store --port 4747 &
+          sleep 5
+          python sft_rollout_runners.py &
+          sleep 5
+          python sft_algorithm.py
+
+          pkill -f agl && echo "SIGTERM sent to agl" || echo "No agl process found"
+          while pgrep -f agl; do
+            echo "Waiting for agl to finish..."
+            sleep 5
+          done
+          pkill -f sft_rollout_runners.py && echo "SIGTERM sent to sft_rollout_runners.py" || echo "No sft_rollout_runners.py process found"
+          while pgrep -f sft_rollout_runners.py; do
+            echo "Waiting for sft_rollout_runners.py to finish..."
+            sleep 5
+          done
+          echo "sft_rollout_runners.py has finished."
+          sleep 10
+
+          # Check models/version_2 must exist
+          if [ ! -d "models/version_2" ]; then
+            echo "models/version_2 does not exist"
+            exit 1
+          fi
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
+
+      - name: Unsloth SFT example all-in-one
+        run: |
+          set -ex
+          source .venv/bin/activate
+          cd examples/unsloth
+          rm -rf models/version_1 models/version_2
+
+          python sft_allinone.py
+          if [ ! -d "models/version_2" ]; then
+            echo "models/version_2 does not exist"
+            exit 1
+          fi
+        env:
+          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
+          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
@@ -0,0 +1,309 @@
+name: Issue Comment
+
+on:
+  issue_comment:
+    types: [created]
+
+permissions:
+  pull-requests: write
+  issues: write
+  contents: write
+  actions: read
+
+jobs:
+  dispatch:
+    # Only run for comments on pull requests AND when the comment starts with "/ci"
+    if: >
+      github.event.issue.pull_request != null &&
+      startsWith(github.event.comment.body, '/ci')
+    runs-on: ubuntu-latest
+    outputs:
+      dispatched: ${{ steps.dispatch.outputs.dispatched }}
+      event_types: ${{ steps.dispatch.outputs.event_types }}
+      correlation_id: ${{ steps.dispatch.outputs.correlation_id }}
+      trigger_comment_id: ${{ steps.dispatch.outputs.trigger_comment_id }}
+      ack_comment_id: ${{ steps.ack.outputs.comment_id }}
+    steps:
+      - name: Guardrail — allow only members/collaborators
+        id: guard
+        uses: actions/github-script@v8
+        with:
+          script: |
+            const allowed = ['MEMBER','OWNER','COLLABORATOR'];
+            const assoc = context.payload.comment.author_association;
+            if (!allowed.includes(assoc)) {
+              core.notice(`Ignoring /ci from ${context.payload.comment.user.login} (author_association=${assoc}).`);
+              core.setOutput('skip', 'true');
+            }
+
+      - name: Trigger repository dispatch
+        id: dispatch
+        if: steps.guard.outputs.skip != 'true'
+        uses: actions/github-script@v8
+        with:
+          script: |
+            const owner = context.repo.owner;
+            const repo = context.repo.repo;
+            const pull_number = context.payload.issue.number;
+            const comment = context.payload.comment;
+
+            // Fetch current PR state
+            const { data: pr } = await github.rest.pulls.get({ owner, repo, pull_number });
+
+            // Add reaction so folks know we saw it
+            try {
+              await github.rest.reactions.createForIssueComment({
+                owner,
+                repo,
+                comment_id: comment.id,
+                content: 'rocket'
+              });
+            } catch (e) {
+              core.info('Could not add reaction (likely due to permissions). Continuing.');
+            }
+
+            const labels = (pr.labels ?? []).map(label => label.name);
+            const directCiLabels = labels.filter(label => label.startsWith('ci-'));
+            const hasCiAll = directCiLabels.includes('ci-all');
+            const dedupe = new Set(
+              directCiLabels.filter(label => label !== 'ci-all')
+            );
+
+            if (!hasCiAll && dedupe.size === 0) {
+              core.notice('No ci-* labels found on the pull request; nothing to dispatch.');
+              core.setOutput('dispatched', 'false');
+              core.setOutput('event_types', '');
+              return;
+            }
+
+            const correlation_id = `id-${comment.id}-${Date.now().toString(36)}`;
+
+            const clientPayload = {
+              correlation_id,
+              pull_number,
+              pr_ref: `refs/pull/${pull_number}/merge`,
+              pr_head_ref: pr.head.ref,
+              pr_head_sha: pr.head.sha,
+              pr_base_ref: pr.base.ref,
+              pr_base_sha: pr.base.sha,
+              trigger_comment_id: comment.id,
+              trigger_comment_user: comment.user.login,
+            };
+
+            const eventTypes = hasCiAll
+              ? ['ci-all']
+              : Array.from(dedupe);
+            for (const eventType of eventTypes) {
+              await github.rest.repos.createDispatchEvent({
+                owner,
+                repo,
+                event_type: eventType,
+                client_payload: { ...clientPayload, ci_label: eventType }
+              });
+              core.notice(`Dispatched '${eventType}' event for PR #${pull_number}.`);
+            }
+
+            core.setOutput('dispatched', 'true');
+            core.setOutput('event_types', eventTypes.join(','));
+            core.setOutput('correlation_id', correlation_id);
+            core.setOutput('trigger_comment_id', String(comment.id));
+
+      - name: Acknowledge in thread (optional)
+        if: steps.guard.outputs.skip != 'true' && steps.dispatch.outputs.dispatched == 'true'
+        id: ack
+        uses: actions/github-script@v8
+        env:
+          EVENT_TYPES: ${{ steps.dispatch.outputs.event_types }}
+          CORRELATION_ID: ${{ steps.dispatch.outputs.correlation_id }}
+        with:
+          script: |
+            const eventTypes = (process.env.EVENT_TYPES || '')
+              .split(',')
+              .map(label => label.trim())
+              .filter(Boolean);
+            const formatted = eventTypes.map(label => `\`repository_dispatch:${label}\``).join(', ');
+            const { owner, repo } = context.repo;
+            const issue_number = context.payload.issue.number;
+            const body = [
+              `✅ CI trigger requested by @${context.payload.comment.user.login}.`,
+              `Fired ${formatted}.`,
+              '',
+              `_Collecting run links for correlation \`${process.env.CORRELATION_ID}\`…_`
+            ].join('\n');
+            const { data: comment } = await github.rest.issues.createComment({
+              owner, repo, issue_number,
+              body
+            });
+            core.setOutput('comment_id', String(comment.id));
+
+      - name: Notify missing ci label
+        if: steps.guard.outputs.skip != 'true' && steps.dispatch.outputs.dispatched != 'true'
+        uses: actions/github-script@v8
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const issue_number = context.payload.issue.number;
+            await github.rest.issues.createComment({
+              owner,
+              repo,
+              issue_number,
+              body: `⚠️ CI trigger ignored because the pull request has no \`ci-*\` labels (e.g. \`ci-apo\`, \`ci-calc-x\`). Add the desired labels and try \`/ci\` again.`
+            });
+
+  watch:
+    needs: dispatch
+    if: needs.dispatch.outputs.dispatched == 'true'
+    runs-on: ubuntu-latest
+    timeout-minutes: 180
+    steps:
+      - name: Track dispatched runs and update comment
+        uses: actions/github-script@v8
+        env:
+          CORRELATION_ID: ${{ needs.dispatch.outputs.correlation_id }}
+          ACK_COMMENT_ID: ${{ needs.dispatch.outputs.ack_comment_id }}
+          TRIGGER_COMMENT_ID: ${{ needs.dispatch.outputs.trigger_comment_id }}
+        with:
+          script: |
+            const owner = context.repo.owner;
+            const repo = context.repo.repo;
+            const correlationId = process.env.CORRELATION_ID;
+            if (!correlationId) {
+              core.warning('No correlation id supplied; nothing to watch.');
+              return;
+            }
+
+            const ackCommentId = Number(process.env.ACK_COMMENT_ID || 0);
+            if (!ackCommentId) {
+              core.warning('No comment id available for updates; skipping watch.');
+              return;
+            }
+            const triggerCommentId = Number(process.env.TRIGGER_COMMENT_ID || 0);
+            if (!triggerCommentId) {
+              core.warning('No trigger comment id available; skipping watch.');
+              return;
+            }
+
+            const prefix = `🚀 CI Watcher for correlation ${correlationId} triggered by comment ${triggerCommentId}`;
+            core.notice(`Watching workflow runs for correlation '${correlationId}' using comment ${ackCommentId}.`);
+
+            function fmt(run) {
+              const status = run.status;
+              const conclusion = run.conclusion;
+              const badge = status === 'completed'
+                ? (conclusion === 'success' ? '🟢' : conclusion === 'failure' ? '🔴' : '🟡')
+                : (status === 'in_progress' ? '🟣' : '⚪️');
+              const title = run.display_title || run.name || `run ${run.id}`;
+              const statusText = status === 'completed' ? `${status}/${conclusion}` : status;
+              return `- ${badge} [${title}](${run.html_url}) — \`${statusText}\``;
+            }
+
+            const signatureOf = runs =>
+              runs
+                .map(run => `${run.id}:${run.status}/${run.conclusion || ''}`)
+                .sort()
+                .join('|');
+
+            const deadlineMs = Date.now() + 175 * 60 * 1000; // 175 minutes
+            let found = [];
+
+            async function searchOnce() {
+              const runs = await github.paginate(
+                github.rest.actions.listWorkflowRunsForRepo,
+                { owner, repo, event: 'repository_dispatch', per_page: 100 }
+              );
+              const cutoff = new Date(Date.now() - 60 * 60 * 1000); // last hour
+              return runs.filter(run => {
+                const createdAt = new Date(run.created_at);
+                const title = String(run.display_title || run.name || '');
+                return createdAt >= cutoff && title.includes(correlationId);
+              });
+            }
+
+            while (Date.now() < deadlineMs) {
+              found = await searchOnce();
+              if (found.length > 0) {
+                core.notice(`Discovered ${found.length} workflow run(s) for correlation '${correlationId}'.`);
+                break;
+              }
+              core.notice(`No runs found yet for correlation '${correlationId}'; retrying shortly.`);
+              await new Promise(res => setTimeout(res, 10000));
+            }
+
+            if (found.length === 0) {
+              core.notice(`Watcher timed out with no runs for correlation '${correlationId}'; notifying thread.`);
+              await github.rest.issues.updateComment({
+                owner,
+                repo,
+                comment_id: ackCommentId,
+                body: [
+                  prefix,
+                  `⚠️ I couldn't find any workflow runs for correlation \`${correlationId}\`.`,
+                  `They may be delayed or misconfigured.`
+                ].join('\n')
+              });
+              return;
+            }
+
+            const runIds = new Set(found.map(run => run.id));
+            let lastSignature = '';
+
+            async function refreshRuns() {
+              const ids = Array.from(runIds);
+              const refreshed = [];
+              for (const id of ids) {
+                const { data } = await github.rest.actions.getWorkflowRun({
+                  owner,
+                  repo,
+                  run_id: id
+                });
+                refreshed.push(data);
+              }
+              return refreshed;
+            }
+
+            async function updateCommentIfChanged(runs, allDone) {
+              const signature = signatureOf(runs);
+              if (signature === lastSignature) {
+                // Run statuses unchanged; skipping comment update.
+                return;
+              }
+              lastSignature = signature;
+              core.notice(`Updating comment ${ackCommentId} with ${runs.length} run status entries (allDone=${allDone}).`);
+              await github.rest.issues.updateComment({
+                owner,
+                repo,
+                comment_id: ackCommentId,
+                body: [
+                  prefix,
+                  `🏃‍♀️ Tracking ${runs.length} workflow run(s):`,
+                  '',
+                  ...runs.map(fmt),
+                  '',
+                  allDone ? '✅ All runs completed.' : '_Still running…_'
+                ].join('\n')
+              });
+            }
+
+            await updateCommentIfChanged(found, found.every(run => run.status === 'completed'));
+
+            while (Date.now() < deadlineMs) {
+              const latest = await searchOnce();
+              for (const run of latest) {
+                if (!runIds.has(run.id)) {
+                  runIds.add(run.id);
+                  core.notice(`Detected additional run ${run.id} (${run.name || run.display_title || 'unnamed'}) for correlation '${correlationId}'.`);
+                }
+              }
+              const current = await refreshRuns();
+              const allDone = current.every(run => run.status === 'completed');
+              await updateCommentIfChanged(current, allDone);
+              if (allDone) {
+                core.notice(`All runs for correlation '${correlationId}' completed; stopping watcher.`);
+                break;
+              }
+              await new Promise(res => setTimeout(res, 60000));
+            }
+
+            if (Date.now() >= deadlineMs) {
+              core.warning(`Watcher hit the deadline while monitoring correlation '${correlationId}'.`);
+            }
@@ -0,0 +1,18 @@
+# Pre-defined workflow with workflow_dispatch trigger,
+# convenient for testing and debugging.
+
+name: Playground
+permissions:
+  contents: read
+on:
+  workflow_dispatch:
+
+jobs:
+  playground:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+      - name: Run script
+        run: |
+          echo "Hello, world!"
@@ -0,0 +1,59 @@
+name: PyPI Nightly Build
+
+on:
+  schedule:
+    # Run daily at 6:00 AM UTC+8
+    - cron: '0 22 * * *'
+  workflow_dispatch:  # Allow manual trigger
+
+jobs:
+  publish-test-pypi:
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write  # IMPORTANT: this permission is mandatory for trusted publishing
+      contents: read
+
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@v6
+        with:
+          python-version: '3.12'
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+      - name: Sync dependencies
+        run: uv sync --frozen --no-default-groups --group dev
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Build dashboard
+        run: cd dashboard && npm run build
+
+      - name: Get current version
+        id: get_version
+        run: |
+          VERSION=$(grep '^version = ' pyproject.toml | sed 's/version = "\(.*\)"/\1/')
+          echo "version=$VERSION" >> $GITHUB_OUTPUT
+          echo "Current version: $VERSION"
+
+      - name: Create development version
+        run: |
+          # Create a dev version with timestamp
+          TIMESTAMP=$(date +%Y%m%d%H%M%S)
+          DEV_VERSION="${{ steps.get_version.outputs.version }}.dev$TIMESTAMP"
+          echo "Creating dev version: $DEV_VERSION"
+          ./scripts/bump_version.sh "$DEV_VERSION"
+
+      - name: Build package
+        run: |
+          uv build
+
+      - name: Publish to Test PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          repository-url: https://test.pypi.org/legacy/
@@ -0,0 +1,81 @@
+name: PyPI Release
+
+on:
+  push:
+    tags:
+      - 'v*'  # Trigger on version tags like v1.0.0, v1.2.3, etc.
+  workflow_dispatch:  # Allow manual trigger
+
+jobs:
+  check-version:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    outputs:
+      version: ${{ steps.get_version.outputs.version }}
+      tag_version: ${{ steps.get_tag.outputs.tag_version }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: Get version from pyproject.toml
+        id: get_version
+        run: |
+          VERSION=$(grep '^version = ' pyproject.toml | sed 's/version = "\(.*\)"/\1/')
+          echo "version=$VERSION" >> $GITHUB_OUTPUT
+          echo "Package version: $VERSION"
+
+      - name: Get tag version
+        id: get_tag
+        run: |
+          TAG_VERSION=${GITHUB_REF#refs/tags/v}
+          echo "tag_version=$TAG_VERSION" >> $GITHUB_OUTPUT
+          echo "Tag version: $TAG_VERSION"
+
+      - name: Verify version matches tag
+        run: |
+          if [ "${{ steps.get_version.outputs.version }}" != "${{ steps.get_tag.outputs.tag_version }}" ]; then
+            echo "Error: Version in pyproject.toml (${{ steps.get_version.outputs.version }}) does not match tag (${{ steps.get_tag.outputs.tag_version }})"
+            exit 1
+          fi
+          echo "Version check passed!"
+
+  publish-pypi:
+    needs: check-version
+    runs-on: ubuntu-latest
+    permissions:
+      id-token: write  # IMPORTANT: this permission is mandatory for trusted publishing
+      contents: read
+
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@v6
+        with:
+          python-version: '3.12'
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+      - name: Sync dependencies
+        run: uv sync --frozen --no-default-groups --group dev
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Build dashboard
+        run: cd dashboard && npm run build
+
+      - name: Build package
+        run: |
+          uv build
+
+      - name: Verify package contents
+        run: |
+          uv run --locked --no-sync python -m tarfile -l dist/*.tar.gz
+          uv run --locked --no-sync python -m zipfile -l dist/*.whl
+
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
@@ -0,0 +1,363 @@
+name: GPU Test
+permissions:
+  contents: read
+on:
+  schedule:
+    # Every day at 5 AM UTC+8
+    - cron: '0 21 * * *'
+
+  workflow_dispatch:
+
+  repository_dispatch:
+    types: [ci-gpu, ci-all]
+
+run-name: >-
+  ${{ github.event_name == 'repository_dispatch'
+      && format(
+        'GPU Test - PR #{0} - {1} - {2}',
+        github.event.client_payload.pull_number,
+        github.event.client_payload.ci_label,
+        github.event.client_payload.correlation_id
+      )
+      || format('GPU Test - {0}', github.event_name) }}
+
+jobs:
+  tests-full:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-gpu' ||
+      github.event.action == 'ci-all'
+    name: Full Test (${{ matrix.mark.display-name }}, ${{ matrix.env.setup-script }}, Python ${{ matrix.env.python-version }})
+
+    runs-on: ${{ matrix.mark.runs-on }}
+    timeout-minutes: 30
+    strategy:
+      matrix:
+        mark:
+          - id: store
+            display-name: Store
+            pytest-mark: 'store'  # store tests should not require gpu
+            runs-on: ubuntu-latest
+            has-gpu: false
+          # AgentOps needs to be separated because it injects tricky global state.
+          - id: agentops
+            display-name: AgentOps
+            pytest-mark: 'agentops'  # including agentops+litellm tests here
+            runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+            has-gpu: true
+          # Similar for Weave.
+          - id: weave
+            display-name: Weave
+            pytest-mark: 'weave'
+            runs-on: ubuntu-latest  # No GPU tests for Weave.
+            has-gpu: false
+          # Other tests that require GPU
+          - id: gpu
+            display-name: GPU required
+            pytest-mark: '(gpu or llmproxy) and not agentops'
+            runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+            has-gpu: true
+          # Other uncovered tests
+          - id: others
+            display-name: Others
+            pytest-mark: 'not store and not agentops and not weave and not gpu and not llmproxy'
+            runs-on: ubuntu-latest
+            has-gpu: false
+        env:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        if: matrix.mark.has-gpu
+        run: nvidia-smi
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.env.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.env.setup-script == 'latest'
+
+      - name: Sync dependencies (latest, gpu)
+        if: matrix.env.setup-script == 'latest' && matrix.mark.has-gpu
+        run: uv sync --frozen --no-default-groups --extra apo --extra weave --extra mongo --group dev --group agents --group langchain --group torch-gpu-stable
+      # Don't install vllm/pytorch on CPU counterparts
+      - name: Sync dependencies (latest, cpu)
+        if: matrix.env.setup-script == 'latest' && !matrix.mark.has-gpu
+        run: uv sync --frozen --no-default-groups --extra apo --extra weave --extra mongo --group dev --group agents --group langchain --group core-stable
+      - name: Sync dependencies (stable, gpu)
+        if: matrix.env.setup-script == 'stable' && matrix.mark.has-gpu
+        run: uv sync --frozen --no-default-groups --extra apo --extra weave --extra mongo --group dev --group agents --group langchain --group torch-gpu-${{ matrix.env.setup-script }}
+      - name: Sync dependencies (stable, cpu)
+        if: matrix.env.setup-script == 'stable' && !matrix.mark.has-gpu
+        run: uv sync --frozen --no-default-groups --extra apo --extra weave --extra mongo --group dev --group agents --group langchain --group core-stable
+      # Don't install langchain for legacy dependency because it has conflicts with torch.
+      - name: Sync dependencies (legacy, gpu)
+        if: matrix.env.setup-script == 'legacy' && matrix.mark.has-gpu
+        run: uv sync --frozen --no-default-groups --extra apo --extra weave --extra mongo --group dev --group agents --group torch-gpu-legacy
+      - name: Sync dependencies (legacy, cpu)
+        if: matrix.env.setup-script == 'legacy' && !matrix.mark.has-gpu
+        run: uv sync --frozen --no-default-groups --extra apo --extra weave --extra mongo --group dev --group agents --group core-legacy
+
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-tests-full-${{ matrix.mark.id }}-${{ matrix.env.python-version }}-${{ matrix.env.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+          cache: 'npm'
+          cache-dependency-path: dashboard/package-lock.json
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Build dashboard
+        run: cd dashboard && npm run build
+
+      - name: Setup Docker environments
+        run: ./scripts/mongodb_docker_run.sh
+        shell: bash
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      # mongo, openai, gpu, all enabled by default
+      - name: Run tests
+        run: |
+          uv run pytest -v --durations=0 tests -m "${{ matrix.mark.pytest-mark }}${{ matrix.env.setup-script == 'legacy' && ' and not langchain' || '' }}"
+        env:
+          PYTEST_ADDOPTS: "--color=yes"
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+          AGL_TEST_MONGO_URI: mongodb://localhost:27017/?replicaSet=rs0
+
+
+  minimal-examples:
+    if: >
+      github.event_name != 'repository_dispatch' ||
+      github.event.action == 'ci-gpu' ||
+      github.event.action == 'ci-all'
+    name: Minimal Examples with Python ${{ matrix.python-version }} (${{ matrix.setup-script }})
+
+    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
+    timeout-minutes: 30
+    strategy:
+      matrix:
+        include:
+        - python-version: '3.10'
+          setup-script: 'legacy'
+        - python-version: '3.12'
+          setup-script: 'stable'
+        - python-version: '3.13'
+          setup-script: 'latest'
+      fail-fast: false
+    steps:
+      - name: Check GPU status
+        run: nvidia-smi
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: uv sync --frozen --no-default-groups --extra apo --group dev --group agents --group langchain --group torch-gpu-stable
+        if: matrix.setup-script == 'latest'
+      - name: Sync dependencies (stable)
+        run: uv sync --frozen --no-default-groups --extra apo --extra mongo --group dev --group agents --group langchain --group torch-gpu-${{ matrix.setup-script }}
+        if: matrix.setup-script == 'stable'
+      # Don't install langchain for legacy dependency because it has conflicts with torch.
+      - name: Sync dependencies (legacy)
+        run: uv sync --frozen --no-default-groups --extra apo --extra mongo --group dev --group agents --group torch-gpu-legacy
+        if: matrix.setup-script == 'legacy'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-minimal-examples-${{ matrix.python-version }}-${{ matrix.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - name: Launch LiteLLM Proxy
+        run: |
+          ./scripts/litellm_run.sh
+        env:
+          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
+          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}
+
+      - name: Write Traces via Otel Tracer
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python write_traces.py otel
+          sleep 5
+
+      - name: Write Traces via AgentOps Tracer
+        env:
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python write_traces.py agentops
+          sleep 5
+
+      - name: Write Traces with Operations
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python write_traces.py operation
+          sleep 5
+
+      - name: Write Traces via Otel Tracer with Client
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          agl store --port 45993 --log-level DEBUG &
+          sleep 5
+          python write_traces.py otel --use-client
+          pkill -f agl && echo "SIGTERM sent to agl" || echo "No agl process found"
+          while pgrep -f agl; do
+            echo "Waiting for agl to finish..."
+            sleep 5
+          done
+
+      - name: Write Traces via AgentOps Tracer with Client
+        env:
+          OPENAI_BASE_URL: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          agl store --port 45993 --log-level DEBUG &
+          sleep 5
+          python write_traces.py agentops --use-client
+          pkill -f agl && echo "SIGTERM sent to agl" || echo "No agl process found"
+          while pgrep -f agl; do
+            echo "Waiting for agl to finish..."
+            sleep 5
+          done
+
+      - name: vLLM Server
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python vllm_server.py Qwen/Qwen2.5-0.5B-Instruct
+
+      - name: LLM Proxy (OpenAI backend)
+        env:
+          OPENAI_API_BASE: http://localhost:12306/
+          OPENAI_API_KEY: dummy
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+
+          python llm_proxy.py openai gpt-4.1-mini &
+
+          LLM_PROXY_READY=0
+          for attempt in $(seq 1 30); do
+            if curl -sSf http://localhost:43886/health > /dev/null 2>&1; then
+              LLM_PROXY_READY=1
+              break
+            fi
+            sleep 2
+          done
+          if [[ "$LLM_PROXY_READY" != "1" ]]; then
+            echo "LLM proxy failed to become healthy" >&2
+            exit 1
+          fi
+
+          python llm_proxy.py test gpt-4.1-mini
+
+          pkill -f llm_proxy.py && echo "SIGTERM sent to llm_proxy.py" || echo "No llm_proxy.py process found"
+          while pgrep -f llm_proxy.py; do
+            echo "Waiting for llm_proxy.py to finish..."
+            sleep 5
+          done
+
+      - name: LLM Proxy (vLLM backend)
+        if: matrix.setup-script != 'legacy'  # Skip if return_token_ids is not supported
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python llm_proxy.py vllm Qwen/Qwen2.5-0.5B-Instruct &
+
+          LLM_PROXY_READY=0
+          for attempt in $(seq 1 30); do
+            if curl -sSf http://localhost:43886/health > /dev/null 2>&1; then
+              LLM_PROXY_READY=1
+              break
+            fi
+            sleep 2
+          done
+          if [[ "$LLM_PROXY_READY" != "1" ]]; then
+            echo "LLM proxy failed to become healthy" >&2
+            exit 1
+          fi
+
+          python llm_proxy.py test Qwen/Qwen2.5-0.5B-Instruct
+
+          pkill -f llm_proxy.py && echo "SIGTERM sent to llm_proxy.py" || echo "No llm_proxy.py process found"
+          while pgrep -f llm_proxy.py; do
+            echo "Waiting for llm_proxy.py to finish..."
+            sleep 5
+          done
+
+      - name: MultiMetrics backend example
+        run: |
+          set -euo pipefail
+          source .venv/bin/activate
+          cd examples/minimal
+          python write_metrics.py --duration 8 --prom-port 9105 --prom-host 0.0.0.0 2>&1 | tee metrics.log &
+          pid=$!
+
+          for attempt in $(seq 1 20); do
+            if curl -sSf http://localhost:9105/metrics | grep -q minimal_requests_total; then
+              echo "Metrics endpoint responding"
+              wait $pid
+              cat metrics.log
+              exit 0
+            fi
+            sleep 1
+          done
+
+          echo "Metrics endpoint did not respond"
+          exit 1
@@ -0,0 +1,238 @@
+name: CPU Test
+
+permissions:
+  contents: read
+
+on:
+  push:
+    branches: [ main, stable/**/* ]
+  pull_request:
+    branches: [ main, stable/**/* ]
+  workflow_dispatch:
+
+  schedule:
+    # Every day at noon and midnight
+    - cron: '0 0,12 * * *'
+
+jobs:
+
+  lint:
+    strategy:
+      matrix:
+        setup: [fast, slow, next]
+      fail-fast: false
+    name: Lint - ${{ matrix.setup }}
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v6
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: '3.12'
+      - name: Sync dependencies (fast)
+        run: uv sync --frozen --group dev --no-default-groups
+        if: matrix.setup == 'fast'
+      - name: Upgrade dependencies (next)
+        run: uv lock --upgrade
+        if: matrix.setup == 'next'
+      - name: Sync dependencies (slow)
+        run: |
+          uv sync --frozen \
+            --extra apo \
+            --extra weave \
+            --extra verl \
+            --extra mongo \
+            --group dev \
+            --group torch-cpu \
+            --group torch-stable \
+            --group trl \
+            --group tinker \
+            --group agents \
+            --group langchain \
+            --no-default-groups
+        if: matrix.setup != 'fast'
+      # This pre-commit skips JavaScript on purpose.
+      - name: Run pre-commit
+        uses: pre-commit/action@v3.0.1
+      - name: Check Python headers
+        run: uv run --locked --no-sync scripts/check_headers.py
+        if: matrix.setup == 'fast'
+      - name: Run Black
+        run: uv run --locked --no-sync black --check .
+        if: matrix.setup != 'next'
+      - name: Run isort
+        run: uv run --locked --no-sync isort --check-only .
+        if: matrix.setup != 'next'
+      - name: Run pyright (fast)
+        run: uv run --locked --no-sync pyright -p pyrightconfig.fast.json
+        if: matrix.setup == 'fast'
+      - name: Run pyright (slow)
+        run: uv run --locked --no-sync pyright -p pyrightconfig.json
+        if: matrix.setup != 'fast'
+
+  lint-js:
+    name: Lint - JavaScript
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v6
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+          cache: 'npm'
+          cache-dependency-path: dashboard/package-lock.json
+      - name: Install dependencies
+        run: cd dashboard && npm ci
+      - name: Run ESLint
+        run: cd dashboard && npm run eslint
+      - name: Run Prettier
+        run: cd dashboard && npm run prettier
+      - name: Run Stylelint
+        run: cd dashboard && npm run stylelint
+      - name: Run Typecheck
+        run: cd dashboard && npm run typecheck
+      - name: Verify build
+        run: cd dashboard && npm run build
+
+  docs:
+    name: Build documentation
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@v6
+        with:
+          python-version: '3.12'
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+      - name: Sync dependencies
+        run: uv sync --frozen --no-default-groups --group dev
+      - name: Set source commit for docs
+        run: |
+          echo "SOURCE_COMMIT=${{ github.sha }}" >> $GITHUB_ENV
+      - name: Verify OpenAPI specification is up-to-date
+        run: |
+          uv run --locked --no-sync python scripts/export_openapi.py
+          git diff --exit-code docs/assets/store-openapi.json
+      - name: Build documentation
+        run: uv run --locked --no-sync mkdocs build --strict
+      - name: Upload docs artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: documentation-site
+          path: site/
+          compression-level: 6
+
+  test:
+    strategy:
+      matrix:
+        mark:
+          # store has many tests and is a good isolated group.
+          - id: store
+            display-name: Store
+            pytest-mark: 'store'
+          # AgentOps needs to be separated because it injects tricky global state.
+          - id: agentops
+            display-name: AgentOps
+            pytest-mark: 'agentops'
+          # Similar for Weave.
+          - id: weave
+            display-name: Weave
+            pytest-mark: 'weave'
+          # litellm proxy tests are slow
+          - id: llmproxy
+            display-name: LLM proxy
+            pytest-mark: 'llmproxy'
+          # Robustness of utilities is important. There are many tests.
+          - id: utils
+            display-name: Utilities
+            pytest-mark: 'utils'
+          # unmarked tests: adapter, execution engine, etc.
+          - id: others
+            display-name: Others
+            pytest-mark: 'not store and not agentops and not weave and not llmproxy and not utils'
+        env:
+          - python-version: '3.10'
+            setup-script: 'legacy'
+          - python-version: '3.11'
+            setup-script: 'stable'
+          - python-version: '3.12'
+            setup-script: 'stable'
+          - python-version: '3.13'
+            setup-script: 'latest'
+      fail-fast: false
+
+    name: Test (${{ matrix.mark.display-name }}, ${{ matrix.env.setup-script }}, Python ${{ matrix.env.python-version }})
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v6
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: ${{ matrix.env.python-version }}
+      - name: Upgrade dependencies (latest)
+        run: uv lock --upgrade
+        if: matrix.env.setup-script == 'latest'
+      - name: Sync dependencies (latest)
+        run: uv sync --frozen --no-default-groups --extra apo --extra weave --group dev --group agents --group langchain --group core-stable
+        if: matrix.env.setup-script == 'latest'
+      - name: Sync dependencies (stable & legacy)
+        run: uv sync --frozen --no-default-groups --extra apo --extra weave --group dev --group agents --group langchain --group core-${{ matrix.env.setup-script }}
+        if: matrix.env.setup-script != 'latest'
+      - name: Freeze dependencies
+        run: |
+          set -ex
+          uv pip freeze | tee requirements-freeze.txt
+          echo "UV_LOCKED=1" >> $GITHUB_ENV
+          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
+      - name: Upload dependencies artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: dependencies-${{ matrix.mark.id }}-${{ matrix.env.python-version }}-${{ matrix.env.setup-script }}
+          path: requirements-freeze.txt
+          compression-level: 0
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+          cache: 'npm'
+          cache-dependency-path: dashboard/package-lock.json
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Build dashboard
+        run: cd dashboard && npm run build
+
+      - name: Run tests
+        run: |
+          uv run pytest -v --durations=0 tests -m "not mongo and not openai and not gpu and (${{ matrix.mark.pytest-mark }})"
+        env:
+          PYTEST_ADDOPTS: "--color=yes"
+
+  test-js:
+    name: Test (JavaScript)
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-node@v6
+        with:
+          node-version: '22'
+          cache: 'npm'
+          cache-dependency-path: dashboard/package-lock.json
+      - uses: astral-sh/setup-uv@v7
+        with:
+          enable-cache: true
+          python-version: '3.12'
+      - name: Sync Python dependencies
+        run: uv sync --frozen --no-default-groups --extra apo --group dev --group agents --group core-stable
+      - name: Install JavaScript dependencies
+        run: cd dashboard && npm ci
+      - name: Run vitest
+        run: cd dashboard && npm run vitest